【作者】 熊丽；

【导师】 钱积新； 梁军；

【作者基本信息】 浙江大学 ， 系统工程， 2008， 博士

【摘要】 随着计算机技术和信息化水平的不断提高,工业企业可以轻易获得大量数据,从不同数据中获取信息加以利用,成就了数据驱动的概念。尽管以多变量统计技术为主体的数据驱动技术通常是建立在一些特定的假设与条件之上,并且存在很多目前尚难以解决的算法与理论上的问题,但多变量统计技术的应用仍然是工厂应用中一个很重要的部分,尤其在计算机技术快速发展的今天,改进技术,运用多种技术实现更多的功能,使得基于数据驱动技术的监测、控制和优化作为工厂应用研究的一个重要方面成为控制学科广泛研究的对象。目前,基于数据驱动的方法和技术在工厂过程各个主要层面的应用都有相关的研究与讨论,如果将各个层面的应用通过数据这个共同的基础联系起来,从底层的回路控制与监测到上层的优化及工艺过程的监控与诊断,便可以通过数据流的连接,在逻辑上形成一个统一的体系,完成数据驱动技术在工厂应用宏观上的统一框架。本文在基于数据驱动技术的工厂控制系统概念及发展状况的同时,对主要的数据驱动技术即多变量统计方法在各个具体应用层面进行了一些分析与研究,提出了一些新的思路,进行了一些具体的仿真与应用。虽然本文无法巨细无遗地对所有问题进行研究,但也希望借助对一些具体层面的探讨抛砖引玉,供作研究的引子与补充。数据驱动技术在工厂过程的应用主要可以分为三大层次:局部回路、局部过程(或批次)、全局过程,针对主要的多变量统计技术在各层次的应用,论文将就一些具体的问题进行研究,主要工作和创新性成果如下:1)介绍基于数据驱动技术的工厂过程控制体系及体系图,并对整体及各个层面进行综述。主要阐述多变量统计技术在建模、多变量统计过程监控与诊断、操作优化及质量优化、控制器设计及控制系统运行和性能的评价、监控与诊断中,从算法到实际应用中需要考虑的各个方面。2)针对主元分析(Principal Component Analysis,PCA)监控无法处理非正态分布采样数据的缺陷,将PCA与核密度估计(Kernel DensityEstimation,KDE)结合及独立元分析(Independent Component Analysis,ICA)与KDE结合,并在一个具体的聚丙烯生产过程应用上对PCA、ICA、KPCA、KICA进行了比较研究和分析。3)针对传统PCA模型线性稳态的局限性,提出一种动态递推PCA(DynamicRecursive PCA,DRPCA)方法,以便对具有动态实时性的系统进行在线监控。结合不同的多变量统计控制图,对两个控制系统(化工蒸馏器、重油分馏塔)及一个三相绕组异步电动机进行了仿真研究,仿真结果证明了方法的有效性。4)石油化工生产的牌号切换过渡过程是运行中最关键、操作难度最大的,具有非常复杂的化学、物理特征和非常敏感的状态突变性其操作成败直接影响到运行的连续性甚至人身、设备的安全。为了提高牌号切换过程运行的平稳性和切换成功率,减少安全隐患,对多向主元分析(DynamicMultiway PCA,DMPCA)及多向偏最小二乘(Dynamic Multiway PLS,DMPLS)进行动态化并对牌号切换过程运行进行监控与诊断,并利用DMPLS对目标性能的预测目标性能指标进行监控。将该方法应用于某石化公司的聚丙烯装置,结果显示该方法可以有效地监控切换状态和产品质量,并对较差的切换过程进行诊断。5)利用PLS隐变量空间主元独立、多变量自动解耦的特点,可以实现通过隐变量自动配对进行控制器设计的方法,但当PLS隐变量之间不能完全消除相关性时,各个回路的PID控制器将不能独立整定,针对这个问题,实行基于动态PLS框架的优化控制,该方法在三进三出化工蒸馏器和重油分馏塔分别进行了仿真研究,结果证明了方法的有效性。6)在牌号切换的操作轨迹优化中引入经验摸型,改变以机理模型为主的情况,用①以PLS模型直接替代机理模型作为约束及②利用PLS逆模型构造优化命题这两种方式将多变量统计回归技术应用于操作优化,在一定程度上简化了建模过程,PLS逆模型也使得优化形式更为简洁。最后,对全文作出总结,归纳了研究和分析的内容及结果,并结合各个方面存在的问题,指出了今后值得关注和进一步深入研究的方向。更多还原

【Abstract】 The concept of "data-driven" is of ten used in computer science. Butbecause of the exceeding progress of computer techniques, massive processdata can be obtained by the intelligentized industry, so "data-driven"has been getting more and more attention in engineering science. Rapidimprovement of capacity and speed of database and massive data obtainedby database make it abstractive how to use them more effectively and tofulfill more functions, but not just show them on panels and screens. Datameans information, and information represents object, so the so-called"data-driven" is drawing information from data, and use information torealize different objects. To draw information from data, statisticaltechniques are the chief methods, and application based on multivariatestatistical techniques become the main part of data-driven area. Atpresent, there have been many papers about different application todifferent level of industry based on data-driven algorithm and techniques.If link the different level application on conjunct data, the wholeindustry process could be covered by control system based on data-driven,that is, data-driven factory. Then from the bottom loop control to upperoptimization to system monitoring and diagnosis, all levels can beconnected by data stream forming a unified logical system. This unifiedsystem can make full use of data in "micro" and "macro" levels. Inthis paper, data-driven industry process control system was broughtforward and its concept and development was discussed as well asapplications to concrete levels were studied. At the same time, severalnew idea and methods were put forward and validated on simulations. Thoughnot all problem could be discussed detailedly here, there is a wish thatthis work could be a complement of study field and could offer a fewcommonplace remarks by way of introduction so that others may come up with valuable opinions.Data-driven industry process control system can be separated as three levels: local loop, local process (or batches) and global process. The main innovational work of this paper are as following:1) Introduce the data-driven industry process control system and the system drawing. Give summarization of the study of the whole system and every level. Mainly expatiate on modeling, multivariate statistical process monitoring and diagnosis (that is, multivariate statistical process control, MSPC), operational optimization and quality optimization, controller design as well as control performance monitoring and diagnosis, based on multivariate statistical techniques. The expatiation is about all aspects including theories, algorithm and application.2) Conventional principal component analysis is based on the assumption that data obeys normal school. If data does not satisfy the assumption, PCA loses its effect, but independent component analysis (ICA) can be more efficient. On the other hand, if data is not normal school, the way calculating control limits will change, and non-parameter method kernel density estimation (KDE) will be used instead of parameter method. So PCA, ICA, PCA with KDE and ICA with KDE were brought forward and applied to a real factory process, the results turned to be effective.3) Principal component model is linear static model, and can not deal with dynamic or time-variant system. To overcome this problem, dynamic recursive algorithm was brought forward, and application to two typical chemical processes and a 3-phase symmetrical induction motors shows effect of the method.4) Consecutive process and batch process monitoring and diagnosis have been studied relatively widely, but trade transition monitoring and diagnosis rarely occur. Only if transition process was included in monitoring and diagnosis system, the system could be complete. Because transition data were intercepted sectional from process, the transition process monitoring and diagnosis could be dealt with by way of batch with MPCA and MPLS. On the other hand, transition processes are of ten nonlinear dynamic process, so MPCA and MPLS were altered to dynamic multiway PCA (DMPCA) and dynamic multiway PLS (DMPLS). DMPLS can also monitor the object quality variables. At last DMPCA and DMPLS were applied to a polypropylene set of a petrochemical company, and the results revealed the capability and potential of the method.5) Some advantages of using PLS as part of control system design include automatic decoupling and efficient loop paring, as well as natural handling of nonsquare systems and poorly conditioned systems. But if the latent variable can not be complete independent, the advantages will disappear and tuning of controller will be difficult. So in this paper, a methodology is proposed for control based on optimization in the subspace defined by the latent variable model, Some simulations were applied to two processes to testify the performance of the method and the results turned out to be affirmative.6) Trade transition is the key part of modern chemical industry, which is very important for the factory benefit. At present, trade transition optimization still mainly use mechanical model, or half mechanical and half experiential model, but rarely use experiential model. While partial least square model (PLS) is experiential model which can make full use of a mass of data, and trade transition operation trajectory optimization using PLS model, both PLS and mechanical model, and PLS reverse latent space were presented which could simplify the modeling step and optimization.Finally, the paper concluded the research findings, and pointed out some future research areas.更多还原